Llama 4 includes several models, primarily the Scout and Maverick, built on a Mixture-of-Experts (MoE) architecture with a significant increase in active parameters and multimodal capabilities. The Scout model is designed to fit on a single H100 GPU, optimizing for lower resource usage while achieving impressive performance on tasks such as coding and reasoning. The Maverick model excels with a larger context window and is geared for multi-GPU systems, affirmatively competing with other leading language models. The Behemoth model is in preview and is expected to outperform notable benchmarks. Users express excitement over the large context window and the model's response capabilities, but some highlight issues with current image generation quality and initial engagements.